Open-weights model trained in an X ads RL environment beats frontier models
kimmonismus · x · 2026-11-10
log(log(ai)) built AdVal, a high-fidelity simulation of X for ad campaigns, and post-trained DeepSeek V4.1 Flash on 181 briefs to create Yvligo, an open-weights model.
Key results
- On 45 held-out briefs, Yvligo scores 77% (geometric mean, relative to best-known conversions per brief), beating Gemini 4 Argon (70%), GPT-6.1 Sol (67%), Grok 4.7 (64%), and Claude Opus 5.5 (63%)
- Live test: $200 budget yielded 383 relevant website visits for Yvligo vs 123 for Grok
- Untuned DeepSeek V4.1 Flash baseline: 64%; rule-based ad group baseline: 59%
The environment, model (Hugging Face), and traces (GitHub) are open-sourced — a striking case of a domain-specific RL environment beating frontier models on a real business task.
More from coding & agent
- Your repo has 5 instruction files but your coding agent only reads 1 — bibryam · 2026-11-10
- OpenAI's dot can now control multiple local computers, picking the right machine per task — OpenAIDevs · 2026-11-10
- Engineer Wraps MCP Calls in LangGraph to Fix Slow, Non-Deterministic Agent — colinmcnamara · 2026-11-10
- 6 patterns behind natural language search: from LLM intent to Typesense queries — bibryam · 2026-11-10
- Dev: Writing Good Comments Is a Harder Problem Than Writing Good Code — neil_conway · 2026-11-10
- OpenAI Agents API ships 12-hour prompt cache, async functions, and per-session budget control — OpenAIDevs · 2026-11-10